Inspiration
Freelancers and small businesses keep invoices and contracts on a home NAS and reconcile them by hand. Matching invoices against contracts is slow and error-prone: overcharges, duplicates, and missed payments slip through, and scattered files make audits painful.
What it does
Invoice–Contract Reconciler is an AI Skill for UGREEN AI NAS. You ask it in plain language — "Reconcile October invoices" — and it finds your invoices and contracts in the folders you authorize, extracts the key terms, matches each invoice to its contract, and writes a dated report listing matched, mismatched, duplicate, overdue, and unmatched items, with the evidence for each finding. After you confirm, it archives everything by vendor and month. It runs entirely on the home NAS, so financial documents never leave home. It never pays an invoice or moves money.
How I built it
The skill is defined in SKILL.md with three reference documents (field glossary, test plan, privacy and safety). It uses three UGOS capability areas — Files, Documents, and Knowledge — orchestrated by four expert agents: a file agent finds the documents, a document agent extracts fields (vendor, number, amount, currency, due date), a reconcile agent matches invoices to contracts and classifies each one, and a report agent writes the dated report and proposes the archive layout. Everything runs on the NAS local AI model; no external services, no credentials, no data transmission. File moves happen only after your explicit confirmation. Note: Round 1 is the skill definition and proposal; installing and tuning it on a real iDX6011 Pro is planned for the Finalist test day.
Challenges
Image-only scans may need OCR — those are flagged for manual review, never guessed. Inconsistent vendor naming lowers match rates. Amounts are compared only in their stated currency. Exact UGOS CLI bindings will be finalized against the Skill Creation Tool during the Finalist test day.
What I learned
Cross-checking two document types against each other only works with strict "never guess" rules: every unreadable field is recorded as missing, every unmatched invoice is listed for review, and nothing is ever deleted or paid automatically.
Disclosure
Muse, an AI agent by Meta, was used to draft this SKILL.md and the accompanying proposal. The skill's workflow, rules, and documentation are original work developed for this Hackathon.
Built With
- ai-agents
- document-processing
- freelancer-tools
- home-server
- invoice-reconciliation
- local-ai
- markdown
- multi-agent
- nas
- on-device-ai
- privacy
- ugos
- ugreen
- ugreen-ai-nas
- workflow-automation
